Real-Time Dense Stereo for Intelligent Vehicles

aroocarmineAI and Robotics

Oct 29, 2013 (3 years and 11 months ago)


Real-Time Dense Stereo for Intelligent Vehicles
Wannes van der Mark and Dariu M.Gavrila
Abstract—Stereo vision is an attractive passive sensing tech-
nique for obtaining three-dimensional (3-D) measurements.Re-
cent hardware advances have given rise to a new class of real-
time dense disparity estimation algorithms.This paper examines
their suitability for intelligent vehicle (IV) applications.In order
to gain a better understanding of the performance and com-
putational cost trade-off,the authors created a framework of
real-time implementations.This consists of different methodical
components based on Single Instruction Multiple Data (SIMD)
Furthermore,the resulting algorithmic variations are com-
pared with other publicly available algorithms.The authors argue
that existing,publicly available stereo data sets are not very
suitable for the IV domain.Therefore,the authors’ evaluation of
stereo algorithms is based on novel realistically looking simulated
data as well as real data from complex urban traffic scenes.In
order to facilitate future benchmarks,all data used in this paper
is made publicly available.
The results from this study reveal that there is a considerable
influence of scene conditions on the performance of all tested
algorithms.Approaches that aim for (global) search optimization
are more affected by this than other approaches.The best
overall performance is achieved by the proposed multiple window
algorithm which uses local matching and a left-right check for
robust error rejection.
Timing results show that the simplest of the proposed SIMD
variants are more than twice as fast than the most complex one.
Nevertheless,the latter still achieve real-time processing speeds
while their average accuracy is at least equal to that of publicly
available non-SIMD algorithms.
Index Terms—Dense disparity,real time,single instruction
multiple data (SIMD),stereo vision.
N appealing application of intelligent transport systems
(ITS) is the automatization of the transport of people and
goods in inner city environments.In order to preserve safety in
such complex environments,current operational systems,such
as people movers,need areas or lanes that are separated from
other traffic.Reliable,robust and real-time obstacle detection
methodologies are needed to enable the safe operation of
these types of intelligent vehicles (IV) among other traffic
participants such as cars and pedestrians.
Stereo vision has the advantage that it is able to obtain an
accurate and detailed 3D representation of the environment
around a vehicle,by passive sensing and at a relatively low
Manuscript received April 8,2005;revised September 9,2005 and October
13,2005.This work was supported in part by the TNO project “Automatised
Safety for Traffic and Transport” (AV3) and by the 5th Framework EU Project
SAVE-U (IST-2001-34040).
W.van der Mark is with the Electro-Optics Group at TNO Defence,
Security and Safety,Oude Waalsdorperweg 63,P.O.Box 96864,2509 JG
The Hague,The Netherlands (
D.M.Gavrila is with the Intelligent Systems Group at the Faculty of
Science,University of Amsterdam,Kruislaan 403,1098 SJ Amsterdam,The
Netherlands ( is also with the Machine
Perception Department of DaimlerChrysler Research,Ulm,Germany.
sensor cost.The work by Labayrade et al.[18] is an example
of a real-time stereo system that is able to detect vehicles up
to 80 m away.This and other previous applications (e.g.[6])
for IV have mostly used sparse,feature-based approaches to
stereo vision.Here only a subset of image pixels (e.g.vertical
edge pixels) are matched,in order to meet real time processing
However,by only using sparse depth data,it is more difficult
to perform a subsequent object segmentation step.For exam-
ple,the vertical edges of a single object are often separated.If
edges of different objects are near to each other it is difficult
to determine which of them belong to the same objects.This
complicates the application of other processing steps such as
classification and tracking because these require some form
of image segmentation.For this reason,it is attractive to use
dense stereo vision,that tries to estimate disparity for all image
A large research community centres around dense stereo
vision because it is attractive for a number of applications such
as robot navigation,surveillance systems,3D modelling,aug-
mented reality and video conferences.Many systems for dense
stereo vision or disparity estimation have been presented,as
discussed in two large surveys of the field,one by Scharstein
and Szeliski [22] and another by Brown et al.[3].
In contrast to previous surveys,this paper does not aim
to review the whole field of dense stereo.Our aim is to
investigate if certain approaches to dense stereo vision are
more suitable for IV applications than others.The criteria
of this investigation are founded on practical considerations
specifically related to the IV domain.
The first of these considerations is that application of dense
stereo in IV is only possible if the disparity map can be cal-
culated in real-time.Single Instruction Multiple Data (SIMD)
offers an appealing and straightforward way for speeding up
computation by carrying out one operation on multiple values
simultaneously.Because the parallelism is only in terms of
the data,difficult problems such as process synchronization
can be avoided.
Over the past few years,manufacturers have extended
general purpose processors with SIMD capabilities in response
to demanding multimedia applications (e.g.SSE2 instruction
set for Intel processors,as used in this paper).Yet SIMD also
forms the basis architecture for special hardware as used in
the ITS domain (e.g.DSPs),which faces particular demands
with respect to power consumption,cost and compactness.
Therefore,in order to evaluate dense stereo vision algorithms
fromIV perspective,it is important to consider their suitability
regarding SIMD parallelism.
In this paper,we identify different methodological compo-
nents and develop efficient underlying SIMD implementations.
The latter are combined in a single framework enabling
comparisons and analysis of the various approaches on an
equal basis.
A second consideration is that the output of stereo algo-
rithms on itself is not interesting for IV applications.Only
subsequent steps,such as obstacle detection or segmentation,
can provide useful information about the vehicle surroundings.
Other work on dense stereo vision has often used error
measures where only the quality of the disparity values was
evaluated.We will present error measures and evaluation
techniques which are more related to typical applications of
stereo vision in the IV domain.
The outline of this paper is as follows.Section II first
discusses a number of concepts for dense disparity com-
putation from the literature.In Section III,we present the
corresponding real-time SIMD implementations.Section IV
compares the resulting algorithms with additional,publicly
available approaches [1],[22] on both simulated and real data
depicting complex urban traffic scenes.Section V contains the
The goal of stereo disparity estimation is finding the correct
correspondences between image points from the left and right
camera.For each point,the positions of possible matches in
the other image is constrained to a single epipolar line,if
the stereo camera geometry is known.Most approaches to
disparity assume that the epipolar lines run parallel to the
image lines,so that corresponding points lie on the same image
lines.This situation can be achieved for stereo cameras by
using a rectification technique [7].If images are rectified,the
disparity d between a corresponding left point l and a right
point r can be expressed as:
d = l ¡r (1)
The disparity space contains all possible matches for the
same left and the right stereo image line.The possible matches
for a point of the left line are a column in this space,the
possible matches for a point on the right line form a row.
Often,a minimum d
and maximum d
disparity are
used to bound this space.Fig.1 shows a drawing of the
disparity space with d
= 0.
Usually,one can distinguish two stages in a disparity estima-
tor.In the first stage,cost values are calculated for comparing
the different points in the disparity search space.These cost
values are used in the second stage for searching the correct
points (matches) in the disparity space.Some algorithms use
additional pre- and post-processing steps.In order to simplify
the matching step,pre-processing is applied to reduce the
illumination differences between the stereo images.A typical
post-processing step is the detection of occlusions,that are
image regions only visible in one of the stereo images.
In the following sections we describe methods from litera-
ture that can be used for each step.
A complicating factor for stereo matching is that the in-
tensities of corresponding pixels from the stereo images can
be different.This can be caused by unequal left and right
camera sensor characteristics such as brightness and contrast.
It is also due to differences in lighting conditions at each of the
camera positions.If raw input images are used,it is necessary
to use an illumination invariant similarity measure.Because
invariant measures are computationally more expensive,pre-
processing of input images is often applied to reduce the
illumination differences beforehand.One approach subtracts
median filtered versions fromthe original input images [24].A
more popular approach is the convolution of the input images
with a Laplacian of Gaussian (LoG) kernel [4].This reduces
illumination influences because the response of the Laplacian
is zero in areas with constant intensity while it is either positive
or negative near edges with high intensity gradient.
Preprocessing can also be used to extract extra information
in order to aid the subsequent disparity search.Hong and Chen
[11] use colour based segmentation to find similarly coloured
patches.They match patches instead of individual pixels be-
cause the assumption is that no large disparity discontinuities
occur within the homogeneous coloured patches.
B.Similarity measures
The simplest similarity measures are based on the difference
in pixel intensity,such as absolute difference (AD) or squared
difference (SD).Algorithms that only use these single intensity
measures in order to compare points are known as “pixel-to-
pixel” algorithms.Unfortunately,discrete images have a quite
limited number of different gray-level intensity values.It is
possible to use colour instead [20].However,colour is difficult
to use during nighttime conditions due to monochromatic
street lightening.
Pixel-to-pixel measures are not very distinctive when inten-
sities of different pixels are the same or corrupted by noise.
Birchfield and Tomasi [1] also pointed out the sensitivity
to image sampling.For example,the pixel intensities on
corresponding stereo edges can be different due to aliasing.
They have therefore designed a measure that is less sensitive
to image sampling.
A more commonly used approach for improving distinction,
is using a larger support region for aggregating the cost values.
Typically,the sum of the pixel differences in a window around
a pixel of interest is used,such as the sum of absolute
differences (SAD) or the sum of squared differences (SSD).
The use of larger windows will lead to more robustness
against noise.However,larger windows with fixed size and
centre point will lead to less accuracy in disparity estimates.
This is due to the fact that areas on slated surfaces will warp
projectively between the stereo images.A square window on
such a surface will therefore only correspond correctly to a
projectively warped version in the other image.
Occlusions near object edges can also cause problems.
When a large window is centred around a background point
near a object edge it will almost certainly encapsulate a portion
of the foreground object.In the case of an occlusion,a large
amount of background pixels in the window will not be visible
in the other image.The resulting similarity measure will
wrongfully be biased towards disparities that belong to the
foreground pixels.
Fig.1.Disparity search space.The
light gray area shows the left-to-right
search area for a point match,while
the dark area shows the right-to-left
search area.
Fig.2.Multiple cost windows.Apart
for the centre window indicated by
the solid lines,four additional win-
dows centred at A,B,C and D are
In the literature,several ways can be found to improve
window based matching.An adaptive size cost window was
proposed by the Kanade and Okutomi [12].Given an initial
guess of disparity,they use a statistical technique in order
to estimate the optimal size and box shape of each matching
window.However,estimating the optimal window size at all
points is computationally expensive.
Other ways of improving the cost computation include
multi-scale techniques and the use of multiple windows.In
some multi-scale approaches several matching windows of
different sizes are used.The larger windows provide the ro-
bustness,while the smaller windows provide precision.On the
downside,matching errors at the coarse scale might propagate
to the finer scales.
A somewhat similar approach is not changing the size of
the windows,but having different options for the location of
the centre point.Both Bobick et al.[2] and Fusiello et al.[8]
use nine different window centre points in their approaches.
For real-time applications,Hirschm
uller et al.[10] suggest
the use of five-window configuration.Fig.2 shows the regular
matching window centred in the middle.On the edge points
(A,B,C,D) of this window four additional windows have been
indicated.Of these four windows the two with the lowest SAD
values are searched.They are added to the value of the centre
window.This step acts as a sort of deformable window,which
means that the resulting window does not have to be centred
on the point itself.
C.Disparity search
When similarity or cost values in the disparity search space
have been calculated,the correct matches can be searched.
A straightforward way of doing this is searching the disparity
interval per point for a optimumvalue.This approach is known
as the Winner-Takes-All (WTA) approach.
A drawback of the simple WTA approach is that it does
not consider the presence of occlusions.This will result in
wrong disparity estimates for those regions.Because WTA
only searches for optimumcost values it is very sensitive to the
results from the cost computation.The cost values computed
for a textureless region often do not indicate a optimum match
due to their similarity.Repetitive texture,such as bars,can lead
to several ambiguous optimums.In these areas,the WTA is
prone to errors.
Assumptions about the scene geometry can be exploited to
improve disparity estimates.Three often applied assumptions
are the “smoothness constraint”,“uniqueness constraint” and
“ordering constraint”.The smoothness constraint is based on
the observation that changes in depth on surfaces are much
smaller than those at the edges of objects.
The uniqueness constraint states that a 3D point only
has exactly one projection in each of the stereo cameras.
Therefore,only one correspondence has to be found for every
stereo point pair.The constraint is only violated by image
points on transparent material or occluded image parts.The
ordering constraint states that the sequence in which points are
ordered on a left image line,is the same for the right image
line.If occlusions are present,points can be missing from
one of the line sequences,but the ordering will remain.There
are scene situations possible where the ordering constraint is
violated [26].However,it is assumed that they are not very
A cost function (CF) is one approach to enforcing smooth-
ness.It can be used to locally add penalties to the values in
the disparity search interval of each next point on the scanline.
Mismatches can be suppressed by penalizing large jumps in
disparity between the scanline points.The difficulty of this
approach is how to choose the penalty function.If penalties
are too high,disparity jumps near edges are missed and if they
are too low the smoothness in not enforced.
Given a correct stereo match in the disparity space,both the
uniqueness and ordering constraint limit the number of pos-
sible matches for the following pixel on the reference image.
Because of the limitations,finding the correct disparities is
now akin to finding a valid path that takes the shortest route
through the cost values.Special rules for how to transverse the
search space can be added in order to handle occlusions and
jumps in disparity.Dynamic-programming (DP) techniques are
often used for this approach [2].
Another search optimization technique called Scanline Op-
timization (SO) was proposed by Scharstein and Szeliski [22].
Instead of enforcing special rules for occlusion and disparity
jumps,they use a global variant of a smoothness cost function
to add penalties to the matching costs.Then a disparity is
searched for each scanline point that minimizes the total cost.
In both the SO and DP the optimization and search tech-
niques are constrained to one single scanline.Just like CF,
they are very dependent on the choice of rules that govern
the path propagation or the cost functions that are used for
searching the optimum solution.This can cause them to miss
or wrongfully suppress large jumps in disparity,for example
near object edges.These types of errors cause “streaks” of
erroneous disparity estimates along scanlines.Interline consis-
tency is difficult to maintain with these techniques.Recently,
some work appeared on optimizing in more than one direction.
The two pass approach of Kim et al.[14],uses the results of
the first optimization pass along the scanlines to optimize the
estimates across the scanlines in the second pass.
An alternative approach is to view disparity search as
a directed graph labelling problem.The graph nodes can
represent left and right pixel pairs with disparity as label.
They can also represent whole pixel patches extracted with
a pre-processing step.Each transition between nodes has a
cost specified by a CF that partially depends on the assigned
disparity label.In several approaches [17],[11],[5] graph cuts
are used to the find disparity label assignments that minimize
the transition costs.The graph optimization approach carries
a high computational cost.However,the algorithms based on
it are among those that currently reside in the top positions of
the performance ranking on the Middlebury stereo vision web
site [28].
This section is devoted to some of the additional steps
performed next to the disparity estimation.These include error
detection,subpixel interpolation and occlusion removal.
It was mentioned earlier that image regions with little
or repetitive texture are problematic for stereo algorithms.
Some approaches try to detect possible errors in the disparity
estimates.Fusiello et al.[8] use the variance of the cost values
in the disparity range of a point as a measure of confidence.
Other approaches [10] select the optimum C
and second best
value C
from the disparity search range.The confidence in
the selected optimum is expressed as:
C =
The idea is based on the fact that the cost values of a
textureless region will be very similar.Repetitive texture will
lead to several optimums in the disparity interval.The measure
of Eq.2 will be low in both cases.However,M
uhlmann et
al.[20] argue that there are occurrences where the optimum
correspondence is actually between two pixels.This also leads
to two very similar optimum values in the disparity search
space.To prevent discarding good estimates on these points,
they suggest using the third best optimuminstead of the second
Besides using integer values for disparity correspondence,
floating-point values can be also used for subpixel accuracy.
Many algorithms use an additional post-processing step for
improving the disparity estimate to subpixel accuracy.If the
SSD similarity measure is used,the cost values near an
optimum can be approximated by a second degree polynomial.
Given the cost values of the optimum and its two nearest
neighbours,the subpixel estimate can be computed by:
= d +
Although this formula is intended for the SSD measure,
many SAD based approaches [10],[20],[25] do also apply it
for subpixel interpolation.
As indicated earlier,an extra step is necessary to detect
and remove erroneous estimates on occlusions by WTA.The
assumptions about scene geometry can also be exploited for
this purpose.
One technique often used is the left-right consistency check.
This check exploits the uniqueness constraint.It is assumed
that the left to right and the right to left disparities are
known.The algorithm checks that the disparity from left
to right correspondence is the same value as when it is
done conversely,searching the match from the right image
to left image.Different value means inconsistency that could
be caused by an occlusion and therefore,these matches are
A drawback of this approach is that the minimum disparity
has to be searched twice for every pixel.Stefano et al.[25]
show that the uniqueness constraint can also be exploited
for occlusion and error removal.During the search for left
to right disparities,the cost value of a best match is stored
for every newly matched right pixel.If a right pixel is
encountered that has been matched before,the new and the
old cost value of the match are compared.The old match
is removed when the new cost value is smaller.This allows
the algorithms to ‘recover’ frompreviously made bad matches.
In order to investigate the suitability of real-time dense
algorithms for the IV domain on an equal basis,we have
implemented our own framework of real-time dense stereo
algorithms.These algorithms use the latest SIMD SSE2 in-
struction set available on the Intel Pentium 4 or AMD Athlon
64 processors.The SSE2 instruction set uses 128-bit registers.
For integer operations these can contain packed 8,16,32 or
64-bit data buffers.
Implementation based on SIMD is not straightforward be-
cause it places restrictions on what kind of operations can
be used.An algorithm can only be sped up if the necessary
operations can be performed in parallel.Conditional constructs
must be avoided because they often lead to stalls of the
code path prediction unit on the processor [9].Furthermore,
processing can only achieve its maximum speed if the values
reside in the (Level 1) cache of the processor.Because of
the much lower speed of main memory,loads and stores to
and from it should be kept to a minimum by doing all the
operations on the data in one go.
These considerations and the real-time requirements for ITS
limit our choice of approaches from the previous section.For
example,the current top ranking algorithms on the Middlebury
stereo vision web site [28],based on the graph cuts technique,
are not suitable due to their computational complexity and
memory requirements.Our implementation therefore consists
of several components,discussed in the previous section,
that are more suitable for real-time operation.They can be
combined to form different stereo algorithms.
The following sections explain how the components in our
framework have been optimized.
A.SAD value computation
The SAD similarity measure can be computed efficiently
by exploiting the fact that neighbouring windows overlap.For
neighbouring windows with the same disparity,the overlap-
ping pixels will contain equal absolute difference (AD) values.
Therefore,a new SAD can be computed out of an old one by
Fig.3.Computation of a new SAD value out of the value of an old one
using intermediate row sums.
subtracting the values,which are only parts of the old window,
and adding the values,which are only parts of the newwindow.
Fig.3 shows how a new SAD value of a window of size
width 2w +1 and height 2h +1 centred at (x;y +1) can be
computed from the previous sum of a window at (x;y).This
process utilizes the intermediate sums of AD values in rows
of the same width as the matching window.Once such a sum
is initialized at the beginning of a line,successive sums can
be calculated recursively by first subtracting the most left AD
value and adding the new AD value at the right-hand side.
Calculated AD values for the additions are stored temporally
in order to re-use them for the subtractions.
Two intermediate sums are applied in order to calculate the
new SAD value of a window at (x;y +1) from the old value
of the window at (x;y).The first intermediate sum is used
for the AD values that belong to the upper row of the old
window.The second sum holds the AD values of the lower
row of the window at the new position.By subtracting the
upper row and adding the lower row,the new SAD value can
be calculated.By repeating this process,a cascade function
is created that only requires two pixel AD’s,four subtractions
and four additions in order to compute a new window sum.
The previously described steps can be executed for multiple
disparity values simultaneously with SIMD type processing.
For the computation of the AD values,SSE2 register xmm0
is filled with 16 copies of the left pixel value.The register
xmm1 holds 16 pixels values from the right disparity interval.
Only two saturated subtractions (xmm0¡xmm1 and xmm1¡
xmm0) and the logical or of the results is needed to calculate
the 16 AD’s.The update of the intermediate sums and the
SAD values themselves are carried out with regular SSE2
subtractions and additions.
B.Multiple windows
As explained earlier,the main deficiency of window based
algorithms is their poor performance near object edges and on
textureless regions.Multiple window approaches can improve
the results.
Our implementation of the five window method by
uller et al.[10] uses a vertical and horizontal step.
In the vertical compare step,newly computed SAD values
are compared with the values of 2h processed image lines
earlier.The results are temporally stored as the minimum
and maximum values for the current ‘centre’ image line of h
image lines earlier.In the horizontal step,the vertical values
on either side (x ¡ w and x + w) of a point are compared.
The maximum of the left hand side is compared with the
minimum of the right hand side and vice versa.The two
resulting minimum values are also the two smallest values
of the four windows.On average,only three compares are
needed to find the smallest values with this method.
C.Left to right minimum search
An important step in a stereo algorithm is searching for the
disparities with the lowest SADvalues.Fig.1 shows the search
space for a disparity interval of 0 · d · max.For a point
l on the left image line,the possible matches with points on
the right image line are indicated with light grey in the search
space.Finding the correct match by searching the match with
the lowest SAD value is computationally expensive because
the whole disparity interval has to be searched.
In [24],a method was shown for finding a minimum using
multimedia extensions (MMX) instruction set that we have
modified to work with SSE2.In this approach two types of
SSE2 registers are used,one holding the range of disparity
values and the other holding a copy of the associated SAD
values.All values of two SAD registers can be compared
in pairs using a single ‘compare smaller than’ instruction.
Both the smaller SAD values and accompanying disparity
values can be selected afterwards,using the resulting binary
mask.This algorithm can be used to process a large disparity
range recursively,after which only eight values remain to be
searched in order to find a single minimum.
D.Right to left minimum search
Disparity search can also be performed from right to left.
Dark gray is used in Fig.1 to indicate possible matches on
the left line for a right point r.
Right to left search does not require recomputation of the
SAD values,the earlier computed left to right values can be
re-used.However,because of the arrangement of the values
in memory,a different algorithm is needed in order to apply
SIMD techniques.By evaluating the SAD value of multiple
right points,instead of one point,this problem can be solved.
Similar to the approach chosen for left to right search,we
start out with an array holding the SAD values and another
holding the disparity values.However,now the SAD values
are compared with the SAD values in the next array.In order
to align the next SAD values,the values are shifted down one
position.By repeating these steps,all the SAD values in the
disparity interval of a right pixel are compared sequentially.
Eventually,the lowest SAD value and its disparity are the ones
discarded from the arrays during the shift step.
E.Search optimizing technique
Almost all real-time SIMD based algorithms known from
literature use the simple WTA technique for disparity search.
We have also implemented a search optimizing method based
on dynamic programming.It is based on a technique,proposed
by Kraft and Jonker [15],that uses two stages;a cost value
propagation stage and a collection stage for retrieving the best
disparity estimates.
Fig.4.Possible predecessor points in cost propagation step.Predecessor
point B has the same disparity as point D,while the points A and C are
respectively an occlusion or part of a discontinuity.
In the cost propagation stage,which runs from left to right
through the disparity space,each point receives a cost value
from a preceding point.The number of preceding points is
limited to three possibilities,that are indicated for a single
point in Fig.4.Each preceding point has a different weight
cost added to its accumulated cost value.Constant weights W
and W
are used for the points A and C which are occlusion
and discontinuity points,respectively.The actual SAD value
of point B is used as its weight because its disparity is equal
to that of the current point.The predecessor with the lowest
total cost value is selected as predecessor of the current point.
Each point also stores the location of its predecessor.These
references link up to form a path trough the disparity search
space.At the end of the propagation stage,the best path is
simply selected by searching for the lowest accumulated cost
value.The best disparities are found in the collection stage by
backtracking this path.
F.Subpixel interpolation
The algorithms of our framework estimate disparity with
subpixel accuracy.Until now,we have described integer based
operations with SSE2.The (older) SSE instruction set also
contains SIMD operations for operations on 128 bit registers
with packed 32 bit floats.We use SSE instructions for comput-
ing Eq.3,which enables subpixel estimation for four points
G.Occlusion removal
We have implemented two types of algorithm components
for removing pixels in occlusions and erroneous matches.The
first type is the left-right check [10],[20],[24].The second
type is the “recover” approach [25].Detected occluded pixels
are set to a predefined error code.This can be a value that is
higher than the disparity maximum or simply zero.
Because of the conditional dependencies used in both ap-
proaches no straightforward application of SIMD is possible.
Fortunately,each check only has to be executed one time
for every pixel.The required overhead is insignificant when
compared to the other steps.
Using the described components we have created seven
different stereo algorithms:
:SAD WTA only left to right search.
:SAD WTA with recover approach.
:SAD WTA with left-right check.
:same as 1,but with multiple windows.
MW5 Rec
:same as 2,but with multiple windows.
:same as 3,but with multiple windows.
:SAD with dynamic programming disparity
search method.
The presented implementations have been tested and com-
pared,together with four other publicly available algorithms.
These are the SSD,DP and SO implementations,created by
Scharstein and Szeliski (S&S) for their survey [22],that are
available on the Internet [28].Furthermore,an implementa-
tion of Birchfield and Tomasi’s (B&T) [1] algorithm in the
OpenCV library [27] is used.
The SSD algorithm is a WTA type algorithm like the
majority of our implementations,however it uses the sum of
squared differences for matching cost computation.Both DP
and B&T use dynamic programming for searching the correct
disparities.In contrast to the other algorithms,B&T does not
use matching windows.Its measure is based on interpolating
values between real pixels to achieve sampling invariance.The
SO algorithm uses scanline optimization for improving the
disparity estimate.
A.Stereo image test sequence
For evaluation purposes,several standard stereo image pairs
with ground truth are available.A well know example are the
stereo pairs of Tsukuba University.Unfortunately,the disparity
range of these pairs is quite small (16 pixels) and the ground
truth disparity is only given with 1 pixel accuracy.
Another well known test set was introduced by Scharstein
and Szeliski in their survey [22].It does provide wider baseline
stereo images with subpixel accurate ground truth disparity.
However,in order to keep the acquisition of ground truth
disparity simple,the images only contain planar surfaces.
Because our goal is to investigate the suitability of dense
stereo vision algorithms for application in the IV domain,
test images are needed which show realistic traffic scenes.
Unfortunately,none of the commonly used test sets resemble
this type of data.
Instead,we use both real and synthetic data of traffic scenes
to evaluate the algorithms.The real image sequences were
recorded with a vehicle mounted stereo camera.These depict
typical traffic scenes with obstacles such as pedestrians and
other cars.Unfortunately,the real images do not come with a
ground truth disparity.It is possible to obtain a ground truth
disparity image with techniques such as active lighting [23].
However,this is not a practical approach for outdoor scenes
and moving stereo camera rigs.Our analysis of the results
with real stereo images is therefore limited to qualitative
Fig.5.Virtual left stereo images from the sequence used for our experiments (frame nr.20,60,100,140,160,220,260 and 300).
Fig.6.Results with simulated stereo images.
Fig.7.Results with real stereo images.
A stereo image pair and its ground truth disparity also
can be generated synthetically from a 3D computer model.
The MARS/PRESCAN software [19],[21] is a framework
for simulation of different vehicle mounted sensors;such
as radar,laser rangefinder or camera based systems such as
stereo vision.With this simulator,a sequence was created of a
virtual vehicle equipped with a stereo camera driving through
traffic scenes.The used city-like scenery is complex and other
moving vehicles are present.
As ground truth,the simulator provides the range images for
each of the stereo images in the sequence.For our experiments,
these are converted to disparity images.
The MARS/PRESCAN synthetic stereo images and ground
truth data used in our experiments is publicly available for
download from the Internet [29].Some of the 326 images of
this sequence,which have a resolution of 512 by 512 pixels
and disparity range of 48 pixels,are shown in Fig.5.
B.Adding real image influences to synthetic images
We first looked at the qualitative similarities between the
output disparity images generated with the simulated images
and real images.The results of the simulated data were very
good for all algorithms,see for example the DP result in the
noiseless case (fig.6b).This is caused by the fact that no
image noise was added to the simulated images.
However,in the output for real stereo images we could more
clearly see errors.An example of a real stereo image of a
vehicle mounted camera is shown in Fig.7a.The output of
dynamic programming approaches such as B&T and DP now
shows a lot of “streaking” errors near object boundaries.WTA
approaches such as in our framework and SSD,on the other
hand,generate mainly errors on areas with insufficient texture.
See for example the road surface in Fig.7c.
In order to approach the conditions of the real images
more closely we added a number of stereo camera related
perturbations.These perturbations are mainly due to the optics,
sensor signal to noise ratio and the calibration of the stereo
rig itself.
Light passing through the edges of a camera lens will hit
the image sensor under a different angle than the light that
passes trough the middle.The light rays at the image edges
are scatted over a larger sensor area than those at the middle.
This effect is known as ‘vignetting’ and causes pixels far away
from the image centre to be darkened.Fig.8,shows the pixel
weights that are multiplied with the original image pixels to
add this effect.The weights where obtained with the cos
which is the technique for calculating vignetting illumination
fall-off [16].
In real cameras two forms of image noise are caused by the
image sensor.The first type is called fixed pattern noise and
is caused by physical differences between the light sensitive
elements on the sensor.However,in almost all modern cameras
the influence of this type of noise is negligible because non-
uniformity correction is used.The other type is called temporal
noise and is due to the sensors signal to the noise ratio.This
type of noise was introduced to the synthetic images by adding
white Gaussian noise with zero mean and a variance of 1
intensity level.
Because dense disparity estimation relies on steps for re-
moving lens distortion and rectifying the stereo images,the
stereo camera calibration itself is also a potential source of
perturbations.In order to add this influence to the “perfectly”
rectified synthetic images,we performed the undistortion and
rectification steps on them with parameters based on typical
residual errors of stereo calibration.Fig.8 shows the mag-
nitudes of the distortion effect between the original and the
distorted synthetic images.
Outputs of the algorithms with these corrupted images now
involve similar artifacts as observed in the real images,e.g.
see SAD
and DP results in Fig.6c & d and Fig.7c
& d.
C.Error measures
Several approaches to quantitative evaluation of stereo al-
gorithms exists.One of the most simplest error measures is
the averaged absolute mean error:
j (4)
Where g is the estimate by the evaluated algorithm and d
is the ground truth.Larger errors can be accentuated by using
the squared error instead of the absolute error:
The drawback of both error measures is that they do not
distinguish well between disparity estimates with a lot of small
errors and disparity estimates with only a few large errors.
Another error measure is the bad pixel percentage.It uses
a threshold ± to set a maximum allowed absolute error.The
absolute differences with the ground truth larger than this value
are counted as bad pixels:
B = 100%
j > ±) (6)
In contrast to the previous two error measures,small errors
are ignored while other errors are counted regardless of their
It is difficult to relate the error measures presented so far
to problems encountered when dense stereo vision is used as
a sensor on a intelligent vehicle.Issues that are of importance
here are the ability to detect objects such as obstacles and
determine their range accurately.
Classifying a group of pixels as an obstacle requires that
they are distinguishable from other background items such as
the road surface,buildings or the sky.We therefore use the
ground truth disparity from the MARS/PRESCAN simulator
to divide pixels in each stereo image into four classes.These
are foreground and background obstacles,road surface and
sky.The two cars are the foreground obstacles in our sequence
while the buildings are background obstacles.Both pixels on
the road and the curb belong to the road surface class.Pixels
that have zero ground truth disparity are classified as being
sky.Examples of the three pixel classes are shown in Fig.9.
A range can only be given for pixels that lie on surfaces,
such as the foreground,background and road pixels.For these
three pixel classes we define the estimation density D of a
disparity image as:
D = 100%
with m=
> 0 (7)
This measures which percentage of the foreground,back-
ground or road surface pixels have been assigned a disparity
For stereo,range has an inverse relationship with disparity;
small disparities correspond to large distances while large
disparities correspond to small distances.Thus,an error in
disparity for a far away point corresponds to a larger error
in range than the same small error in disparity for a nearby
point.The error measures of Eq.4,5 and 6 do not take this
into account.In the Mean Relative Error measure the absolute
error is divided by the ground truth disparity for each pixel.
Therefore,this measure does actually relate to the expected
error in range estimation.
> 0)

Fig.8.Influences on the quality of a stereo image pair.
For ITS applications it might be perceived that only the
performance measures for the foreground obstacles are im-
portant.However,obstacle detection itself involves finding
the correct foreground pixels among the other pixels classes.
Since the road surface is more easily distinguishable than the
(smaller) foreground objects it is actually searched first in
some approaches.The performance measures for the other
classes are therefore significant because bad estimates here
can cause false positives.
D.Algorithms of the framework
The SIMD based algorithms of our framework and others
from the public domain were tested with our sequence.The
window based algorithms all used the same square window
size of 9 by 9 pixels for cost computation.The weights for
were set so that W
= 34000 and W
= 1000.The
tests were conducted on an Intel Pentium 4 3.2 GHz PC with
1.0 GB RAM.
The results for the full resolution frames 1 until 90 are
shown in Table I.For each of the surface pixel classes the
rejection percentage R%,estimate density D% and averaged
relative mean error E
are shown.The fourth column shows
the overall results averaged over all the surface pixels.
We first studied the overall performance of the different
algorithms from our framework.The algorithms that do not
reject pixels show the highest error rates,see Table I.These are
mainly caused by erroneous estimations in difficult areas such
as occlusions and textureless regions.The algorithms with
post-processing have lower error rates because they manage
to reject many of these pixels.Comparing the percentages of
pixels rejected by the recovery and the left-right approach,it
is clear that the left-right approach rejects more pixels.Our
DP approach rejects the least amount of pixels.
Considering the timing results for both half and full res-
olution frames in Table II,it is clear that the improvement
by the post-processing steps comes at the price of a higher
computational cost.The left-right consistency check is more
expensive than the recovery approach.If the processing times
of SAD
and SAD
algorithms are compared to the
time needed for SAD
,the recovery approach shows a 10%
increase,while the left-right check shows a 20% increase.
The DP approach has the highest computational cost of our
optimized algorithms.This is due to the fact that references
have to be stored for back propagation phase,which increases
the number of expensive memory operations.
It should be noted that the achieved run times only give
an outlook on future performance because they have been
achieved on general purpose computer hardware.For ITS
it is much more likely that more dedicated and low-power
embedded SIMD hardware will be used.
As follows from Table I,the multiple window approach
does improve results.It also decreases the number of rejected
pixels.Of all our algorithms the SAD
algorithm has
the lowest mean relative error for all surfaces pixels.From the
timing results however,it is clear that the multiple window
approach is expensive compared to the single window ap-
proach.The multiple window approach increases computation
by about 50%.
Fig.9.The different pixel classes used for evaluation.
Fig.10.The estimation densities of six algorithms on foreground pixels.
Fig.11.The mean relative error of six algorithms on foreground pixels.
E.Performance for different surface pixel types
The previous section provided an insight how well our vari-
ous algorithms compare to each other and what computational
cost they incur.In this section we investigate the performance
figures for separate pixel classes.We have also included the
results of other,publicly available algorithms in this analysis.
The relative mean error for all surface pixels in Table I
shows that the SAD
has the lowest overall error rate,
followed by S&S DP as second best.However,this is not the
case for all individual surface pixel classes.
For the background and road pixel classes it can be seen that
approaches based on dynamic programming such as SAD
and S&S DP perform better than many of the WTA based
algorithms.This is due to the fact that the road and background
classes do not contain many disparity discontinuities,their
disparity profiles adhere to the smoothness constraint.Because
disparity discontinuities do occur at the edges of the fore-
ground objects,the DP based approaches have more problems
with these type of pixels.The WTA based algorithms that use
the left-right check to remove errors are more successful on
foreground pixels.
Regarding the processing time of various algorithms,Table
II reveals the benefit of the pursued SIMD SSE2 implementa-
tion.Our optimized algorithms are much faster than the SIMD
non-optimized algorithms S&S DP,S&S SO,S&S SSD and
B&T DP.The B&T DP is much faster than the S&S DP
algorithm,because although both use dynamic programming,
it is based on a pixel-based sampling invariant similarity
measure,rather than the window-based measure used by S&S
F.Analysis of performance variations
Until now,we have only considered averaged results for a
part of the sequence.However,the performance of the tested
algorithms clearly varies from frame to frame.This can seen
in plots of the estimation density percentage and relative mean
error per stereo image pair.
In Fig.10 two plots are shown of estimation density
percentages (D%) on foreground pixels for frame 0 until 90.
The plot on left side shows the results of the WTA
and WTA
algorithms from our framework
while the plot of the right side shows the results of S&S SO,
S&S DP and B&T DP.Fig.11 shows two plots for the same
interval.This time,the relative mean error (E
) is shown for
the six algorithms.
During this part of the sequence,a foreground object (a car)
passes a road crossing.The vehicle drives into in the camera
field of view from a side street.It is completely visible in
frame 25.After frame 48 it starts to leave the cameras field
of view.It is completely out of the field of view after frame
If the different density percentage plots are compared it is
clear that the algorithms from our frame work show a gradual
decrease.This is due to the fact that the surface of the car
contains little texture.In the beginning,when the car is far
away,features such as edges,provide enough distinct points
for the stereo matching.When the car is nearby more pixels
in textureless areas are visible.This causes the decrease in
estimation density.
Because WTA
,S&S SO,S&S DP and B&T DP use
search optimization techniques they are less effected by
textureless regions.The plot of the latter three algorithms
does show lower densities percentages after frame 48,which
due to ‘streaking errors’.Because the disparity jump from
background to foreground disparity is not visible anymore after
frame 48,the dynamic programming technique wrongfully
assigns zero disparities to some of the foreground pixels.
The effects of streaking errors by algorithms with search
optimization is more evident in the plots of Fig.11 where the
relative mean error is shown.The algorithms WTA
that do not use search optimization techniques
show a fairly consistent error of about 0.2,while the other
algorithms,that do use DP or SO show a sharp increase in
error after frame 48.
G.Ground plane estimation experiment
In the previous sections we have studied the quantitative
results of the different disparity estimators.We learned that
the algorithms which use global search techniques are more
effected by the scene complexity.In this section we will show
some of the consequences for a typical application of stereo
vision in intelligent vehicles:ground plane estimation.
Ground plane estimation is required to distinguish obstacles
such as other cars and pedestrians in the disparity image from
the road surface.A robust and real-time method for doing this
was developed by Labayrade et al.[18].
It is based on the assumption that for scanlines where the
road surface is visible,the dominant disparity value is that
of road surface pixels.Their method first converts the normal
disparity image to a ‘V-disparity’ image.Each scanline of such
an image is the histogramof disparity values of a scanline from
the original disparity image.The road surface profile can be
extracted from the ‘V-disparity’ image by finding dominant
line features.The original method of Labayrade et al.[18] uses
the Hough transformto find line features and approximates the
road vertical curvature with a piecewise linear curve.
For our experiment,we use a simplified version of the
Labayrade et al.[18] approach.To test an estimated disparity
image it is first converted to a V-disparity image.Because
the road surface of the synthetic images is flat,only a single
dominant line feature is searched with the Hough transform.
This line is then compared to the line found by the same
method in the ground truth disparity image.The difference
in angle between the two lines shows how ground plane
estimation is affected by the quality of the disparity image.
In Fig.12 the differences in ground plane angle is shown for
all images from test sequence.Each column shows the errors
colour coded in degrees for one of the tested algorithms.
Fig.12.Error in ground plane angle estimation based on V-disparity.
These results show that error made by algorithms,which use
optimization steps,affect the ground plane estimation more
than the more simpler approaches.The SAD
,S&S DP,
S&S SO and B&T DP show severe errors in ground plane
angle estimation in some the frames.Both the SAD
and S&S SSD also show severe errors because they do not use
rejection steps.It is also interesting to see that the multiple
window approach,when compared to the single window
approach,actually increases the error slightly for some parts of
the sequence.This is due to the fact that V-disparity assumes
that a majority of disparities belong to the road surface.It
can therefore become biased in situations where this is not
true.The multiple window approach gives more estimates
for the textureless surface of the passing car than the single
window approach.Surprisingly,this influences the ground
plane estimation when the car is nearby and fills a large part
of the image.
Application of dense stereo vision in intelligent vehicles
requires accurate and robust disparity estimation algorithms
that can run in real-time on small and power efficient com-
puting hardware.Dense stereo vision algorithms which are
based on Single Instruction Multiple Data processing are
interesting because this type of instruction level parallelism is
currently available on various normal,low power embedded
and dedicated computing systems.
We implemented several real-time algorithms to investigate
how well different approaches to dense stereo vision can
benefit from SIMD optimization and compare them on an
equal basis.They are all based on a single framework that
provides components for performing SAD cost computation,
multiple window selection,disparity search based on WTA or
dynamic programming and post-processing for occlusion or
error rejection.We came up with fast SIMD optimized ver-
sions for most of these components using the SSE2 instruction
In order to test the algorithms under the varying conditions,
that can be encountered in city-like traffic,we used stereo
images from a vehicle sensor simulator.In contrast to the
commonly used single pair test images,our sequence con-
tains hundreds of images with varying geometric and image
properties.Effects that degrade image quality in real stereo
cameras,such as vignetting,sensor noise and imperfect image
undistortion and rectification were also added to enhance
The first set of tests show that the post-processing steps
can help reduce error percentages at the cost of a sparser dis-
parity map.More advanced algorithms with multiple window
technique or dynamic programming improve the disparity es-
timates on difficult image areas,albeit at higher computational
cost.Nevertheless,the slowest of our algorithms still achieves
real-time processing speeds.
We also studied the performance of our algorithms together
with a set of four other publicly available stereo algorithms
for different types of foreground disparity pixel classes.The
results confirm that non-greedy matching algorithms (i.e.
scan line optimization,dynamic programming) perform better
on surfaces with low geometrical complexity such as the
road surface.However,the more simpler WTA techniques
combined with error detection techniques can outperform
these algorithms on more challenging surfaces such as nearby
obstacles.This is caused by the fact that search optimization
algorithms can miss important disparity jumps at the edges of
nearby objects and therefore assume the wrong disparities in
later search or optimization stages.
The consequences of these type of errors are demonstrated
with an road surface inclination estimation experiment.The
results show that ground plane estimation based on disparity
estimates of the algorithms which do use optimization tech-
niques are less reliable.
Our research has shown that simple WTA techniques for
dense stereo,combined with robust error rejection schemes
such as the left-right check,are more suitable for intelligent ve-
hicle applications.Because they do not use search optimization
techniques their processing speed is higher.Our investigation
has also revealed that the basic search optimization techniques
such as dynamic programming can cause errors which interfere
with subsequent steps needed for intelligent vehicle related
sensing tasks.
The results suggest that future improvement of dense stereo
algorithms can be achieved by applying search optimization
techniques only to the parts of the stereo images that can
benefit from them,as discussed above.Preprocessing steps
to detect textureless and edgeless regions could be exploited
to detect which pixels are suitable.
The authors would line to thank M.G.van Elk from
the Imaging Systems (IST) Signal Processing Department at
TNO Science and Industry for generating the sequence of
synthetic stereo images and ground truth range data used in
our experiments.The authors also thank Harris Sunyoto for
his assistance in performing the early-stage experiments.
S.Birchfield,C.Tomasi,“Depth Discontinuities by Pixel-to-Pixel Stereo”,
International Conference on Computer Vision,pp.1073-1080,Bombay,
A.F.Bobick,S.S.Intille,“Large occlusion stereo”,International Journal
of Computer Vision,Vol.33,Issue 3,pp.181-200,1999.
M.Z.Brown,D.Burschka,G.D.Hager,“Advances in Computational
Stereo”,IEEE Transactions on Pattern Analysis and Machine Intelligence,
25(8),August 2003.
ose,“Range Estimation from a Pair of Omni-
directional Images”,IEEE International Conference on Robotics and
H.Y.Deng,Q.Yang,X.Lin,X.Tang,“A Symmetric Patch-Based Corre-
spondence Model for Occlusion Handling”,To appear in Proceedings
IEEE International Conference on Computer Vision,Beijing,China,
October 15-21,2005.
U.Franke,S.Heinrich,“Fast Obstacle Detection for Urban Traffic
Situations”,IEEE Transactions on Intelligent Transportation Systems,
Vol.3,No.3,September 2002.
A.Fusiello,E.Trucco,A.Verri,“Recification with unconstraint stereo
gemometry”,Proceedings of the Eighth British Machine Vision Confer-
A.Fusiello,V.Roberto,and E.Trucco.“Symmetric stereo with multiple
windowing”.International Journal of Pattern Recognition and Artificial
Intelligence,14(8) pp.1053-1066,December 2000.
R.Gerber,The Software Optimization Cookbook,Intel Press,Intel Coop-
uller,P.R Innocent,J.M.Garibaldi,“Real-Time Correlation-
Based Stereo Vision with Reduced Border Errors”,International Journal
of Computer Vision,Vol.47(1/2/3),pp.229-246,2002.
L.Hong,G.Chen,“Segment-based stereo matching using graph cuts”,
Proceedings IEEE Conference on Computer Vision and Pattern Recogni-
tion,Vol.I,pp.74-81,Washington,DC,USA,27 June - 2 July,2004.
T.Kanade,M.Okutomi,“A stereo matching algorithm with an adative
window:theory and experiment”,IEEE Transactions on Pattern Analysis
and Machine Intelligence,Vol.16,Issue 9,pp 920-932,September 1994.
T.Kanade,H.Kano,S.Kimura,“Development of a VideoRate Stereo
Machine”.International Robotics and System Conferences,pp.95-100,
August 1995.
J.C.Kim,K.M.Lee,B.T.Choi,S.U.Lee,“A Dense Stereo Matching
Using Two-Pass Dynamic Programming with Generalized Ground Con-
trol Points”,Proceedings IEEE International Conference on Computer
Vision and Pattern Recognition,Vol.II,pp.1075-1082,San Diego,CA,
USA,June 20-25,2005.
G.Kraft,P.P.Jonker,“Real-Time Stereo with Dense Output by a SIMD-
Computed Dynamic Programming Algorithm”,International Conference
on Parallel and Distributed Processing Techniques and Applications,Vol.
III,pp.1031-1036,Las Vegas,Nevada,USA,June 24-27,2002.
C.Kolb,D.Mitchell,P.Hanrahan,“A Realistic Camera Model for
Computer Graphics”,Computer Graphics (Proceedings of SIGGRAPH
’95),ACM SIGGRAPH,pp.317-324,1995.
V.Kolmogorov,R.Zabih,“Computing Visual Correspondence with Oc-
clusions using Graph Cuts”,Proceedings IEEE International Conference
on Computer Vision (ICCV),Vancouver,Canada,July 9-12,2001.
R.D.Labayrade,J.P.Tarel,“Real Time Obstacle Detection on Non Flat
Road Geometry through ‘V-Disparity’ Representation”,Proceedings of
IEEE Intelligent Vehicle Symposium,Versailles,France,18-20 June 2002.
F.J.W.Leneman,“An integrated design and validation environment for
intelligent vehicle safety systems (IVSS)”.10th World Congress and
exhibition on ITS,Proceedings on CD-ROM,Madrid,Spain,16-20 Nov
anner,“Calculating Dense
Disparity Maps from Color Stereo Images,an Efficient Implementation”,
International Journal of Computer Vision,Vol.47,Nr.1-3,pp.79-88,
April - June 2002.
Z.Papp,K.Labibes,A.H.C.Thean,M.G.van Elk,“Multi-Agent Based
HIL Simulator with High Fidelity Virtual Sensors”,IEEE Intelligent
Vehicles Symposium,pp.213-218,Columbus (OH),June 9-11,2003.
D.Scharstein,R.Szeliski,“A Taxonomy and Evaluation of Dense
Two-Frame Stereo Correspondence Algorithms”,International Journal
of Computer Vision,Vol.47 (April-June),pp.7-42,2002.
D.Scharstein,R.Szeliski,“High-accuracy stereo depth maps using
structured light”.IEEE Computer Society Conference on Computer Vision
and Pattern Recognition,Vol.1,pp.195-202,Madison,WI,June 2003.
L.Di Stefano,S.Mattoccia,“Fast Stereo Matching for the VIDET
System using a General Purpose Processor with Multimedia Extensions”,
Fifth IEEE International Workshop on Computer Architectures for Ma-
chine Perception,pp.356-362,Padova,Italy,September 11 - 13,2000.
L.Di Stefano,M.Marchionni,S.Mattoccia,G.Neri,“A Fast Area-
Based Stereo Matching Algorithm”,15th IAPR/CIPRS International
Conference on Vision Interface,Calgary,Canada,May 27-29,2002.
M.Ziegler,Region-based analysis and coding of stereoscopic video,
PhD Thesis,Technische Universiteit Delft,The Netherlands,1997.
The Open Source Computer Vision Library,available online:
Middlebury College Stereo Vision Research Page,available online:
Stereo Image Data for Algorithm Evaluation,available online:
Wannes van der Mark was born in Leiderdorp,
the Netherlands,on 22 June,1975.He obtained
the M.Sc.Degree in Artificial Intelligence from
the University of Amsterdam in 2000.He currently
works as a PhD student at both TNO Defence,
Security and Safety in The Hague and the University
of Amsterdam.His current research interest is in
stereovision for autonomous vehicle guidance in
unstructured terrain.
Dariu M.Gavrila obtained the M.Sc.Degree in
Computer Science from the Free University in Am-
sterdam in 1990.He received the Ph.D.Degree in
Computer Science from the University of Maryland
at College Park in 1996.He was a Visiting Re-
searcher at the MIT Media Laboratory in 1996.Since
1997 he is a Research Scientist at DaimlerChrysler
Research in Ulm,Germany.In 2003,he was ap-
pointed Professor at the University of Amsterdam,
chairing the area of Intelligent Perception Systems
Mr.Gavrila’s long-term research interests involve vision systems for
detecting human presence and activity with applications in intelligent vehicles
and surveillance,in which he has numerous publications.His personal website
is “”.